How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

string-diagrams-lean - Skill

Name: string-diagrams-lean
Author: apellis

Imported from apellis/string-diagrams-lean (AGENTS.md). Install upstream with npx skills add apellis/string-diagrams-lean. Copyright stays with the author.

Contributor guide

This repository is developed largely by AI agents under human direction. Everything here is public: keep it mathematical and technical.

Ground rules

The toolchain (lean-toolchain) and Mathlib and sibling-library pins (lake-manifest.json) are fixed. Do not run lake update except as a deliberate, coordinated bump. Build with lake exe cache get and then lake --wfail build. Warnings are errors.
Before presenting a result as done, run lake env lean -DwarningAsError=true scripts/AxiomAudit.lean (CI runs it as well). Committed files contain no sorry, admit, native_decide or new axiom. The axiom closure is contained in propext, Classical.choice and Quot.sound. Compilation alone is not the acceptance criterion; source fidelity is.
Dependencies: Mathlib only. This library is generic. It must not import OddMath or Categorification, and odd-specific generators, signs and interpretations belong downstream.
Record the source of every main result in its docstring: paper and version, section, and theorem or equation number. Check each locator in the cited version. Distinguish the literal printed statement, a repaired statement, a proof gap and a counterexample. Never weaken a hypothesis or change a sign silently to make something compile.
New modules go in a topical directory under StringDiagrams/, with a module docstring that names the source results they formalize.
Commit messages describe the mathematics, not the process.

Staying on course

These rules apply to every contributor and take precedence over local habits or private instructions that conflict with them.

Targets and scope

Every PR advances a specific numbered result of a cited source (theorem, proposition, corollary, lemma, definition or equation) or a roadmap item that one of them needs. The PR title or body names it. Supporting lemmas land with, or shortly before, the result that uses them.
Formalize what the source states. Do not label results "original", and do not develop new theory when a source step is missing. When a cited proof relies on a reference that is not available, on a step that has a gap, or on a result outside the library, stop and record this in the tracking issue as a question for the maintainer. The options are: (a) obtain the reference; (b) take the missing input as an explicit, documented hypothesis of the dependent results; (c) build a replacement route. Do not start (c) without the maintainer's approval.
Scope decisions are the maintainer's. These include declaring a numbered result out of scope, choosing between several valid conventions, and substituting a different object for the one the source uses (for example a crystal basis for a canonical basis). List them as open questions; do not decide them silently.
Prove one form of a result. Restate it in another form (maps → bimodule maps → natural isomorphisms → mates, and so on) only when a named downstream result needs that form.

Statements

A headline statement carries exactly the source's hypotheses, after any recorded erratum correction.
An extra hypothesis that the current proof route needs is allowed only on an intermediate statement. Its docstring must say so, the README status must say so, and a roadmap item must track its removal.
Names, docstrings and status tables never overstate. Do not use _unconditional on a conditional theorem, and do not use "full" or "complete" for a statement that assumes part of its conclusion.
An erratum entry gives the printed statement, the corrected statement with the extra hypothesis or correction, the declarations, and a short paragraph that can be sent to the authors as is.

Lean engineering

Elaboration trouble is not a mathematical obstruction. Examples: whnf or isDefEq timeouts on dependent transports, heartbeat limits, motive errors, slow instance search.
After two failed attempts at the same proof, redesign the encoding. Possible changes: reindex so that the types agree definitionally, prove one generic lemma instead of many instances, or avoid nested casts. Do not raise maxHeartbeats or retry with a larger budget.
Names of declarations, modules and PRs describe the mathematics. Avoid process adjectives ("actual", "genuine", "original", "native", "fixed", "frontier"), stage labels (LegA, AdapterB7, PbwL3) and "WIP". A PR whose title says WIP is not merged.

Workflow and evidence

All proof work happens on topic branches of this repository and lands through PRs with green CI. Work that does not build yet may be pushed as a draft PR. Proof developments are not kept in private copies, scratch archives or other repositories.
The evidence that a result is done is CI: full build with warnings as errors, plus the axiom audit. Agreement between reviewers, private sign-offs, hash manifests and log archives are not evidence. Do not commit logs, manifests, archives or review packets.
The same PR updates the README/ROADMAP status row and the tracking issue. Close an issue when its goal is met.

Install string-diagrams-lean and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

{ "ocm": "1", "id": "apellis-string-diagrams-lean-string-diagrams-lean", "kind": "skill", "name": "string-diagrams-lean", "description": "This repository is developed largely by AI agents under human direction. Everything here is public: keep it mathematical and technical.", "publisher": "apellis", "version": "1.0.0", "capabilities": { "domains": [ "math", "coding" ], "tags": [ "skill-md", "github-agents" ], "languages": [ "en" ] }, "quality_prior": 0.6, "examples": [ "This repository is developed largely by AI agents under human direction. Everything here is public: keep it mathematical and technical." ], "primary": false, "metadata": { "source": { "provider": "github-agents", "repository": "https://github.com/apellis/string-diagrams-lean", "path": "AGENTS.md", "ref": "9976974dfbce9c140f20c41467abbe44a4a0e294", "url": "https://github.com/apellis/string-diagrams-lean/blob/9976974dfbce9c140f20c41467abbe44a4a0e294/AGENTS.md", "key": "apellis/string-diagrams-lean/AGENTS.md" } }, "instructions": "# Contributor guide\n\nThis repository is developed largely by AI agents under human direction. Everything here is\npublic: keep it mathematical and technical.\n\n## Ground rules\n\n- The toolchain (`lean-toolchain`) and Mathlib and sibling-library pins (`lake-manifest.json`)\n are fixed. Do not run `lake update` except as a deliberate, coordinated bump. Build with\n `lake exe cache get` and then `lake --wfail build`. Warnings are errors.\n- Before presenting a result as done, run `lake env lean -DwarningAsError=true scripts/AxiomAudit.lean`\n (CI runs it as well). Committed files contain no `sorry`, ", "cost": { "context_tokens": 1288 } }

string-diagrams-lean

About

Contributor guide

Ground rules

Staying on course

Install string-diagrams-lean and it becomes one more candidate the router can pick - when it fits.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Use it

Manifest

Reviews